The official implementatioin of paper "UD-V2X: Uncertainty-driven Vehicle-to-everything Collaborative Perception".
Collaborative perception enables each agent to achieve a more comprehensive and robust awareness of the surrounding environment over individual perception. Although the end-to-end approach significantly enhances the perception performance, previous collaboration methods often overlook the crucial role of uncertainty in individual perception results. We propose leveraging the uncertainty in individual perception results to guide communication and fusion in the collaborative process, which offers interpretability and varying levels of attention. In this paper, we propose UD-V2X, a novel multi-agent collaborative perception framework with three key components: i) a learning-based uncertainty generator, which can estimate the uncertainty of individual perception results at a minimal computational cost; ii) an uncertainty-driven spatial feature selection, which could not only address the missed detection issue but also mitigate false positive problems that previous objectiveness-based communication methods could not resolve; and iii) an uncertainty-aware enhanced mechanism (UEM), which is a plug-and-play module that can be seamlessly integrated into existing collaborative perception frameworks to refine the extracted feature maps. We evaluate UD-V2X on three large-scale open-source benchmarks, including the simulated datasets OPV2V, V2XSet and the real-world dataset DAIR-V2X. The experimental results demonstrate that our UD-V2X achieves superior performance compared to the previous state-of-the-art methods. Our code is available at https://github.com/uestchjw/UD-V2X.
Please visit CoAlign for details.
Since our work has not been accepted yet, please contact us ([email protected]) if you would like a checkpoint.